WO2018160943A1 - Procédé et appareil de détection d'états d'arnaque - Google Patents

Procédé et appareil de détection d'états d'arnaque Download PDF

Info

Publication number
WO2018160943A1
WO2018160943A1 PCT/US2018/020624 US2018020624W WO2018160943A1 WO 2018160943 A1 WO2018160943 A1 WO 2018160943A1 US 2018020624 W US2018020624 W US 2018020624W WO 2018160943 A1 WO2018160943 A1 WO 2018160943A1
Authority
WO
WIPO (PCT)
Prior art keywords
features
dnn
deep
acoustic features
voice
Prior art date
Application number
PCT/US2018/020624
Other languages
English (en)
Inventor
Elie Khoury
Parav NAGARSHETH
Kailash PATIL
Matthew Garland
Original Assignee
Pindrop Security, Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Pindrop Security, Inc. filed Critical Pindrop Security, Inc.
Priority to AU2018226844A priority Critical patent/AU2018226844B2/en
Priority to CA3054063A priority patent/CA3054063A1/fr
Priority to EP18712052.2A priority patent/EP3590113B1/fr
Publication of WO2018160943A1 publication Critical patent/WO2018160943A1/fr
Priority to AU2021277642A priority patent/AU2021277642B2/en

Links

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/02—Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L17/00—Speaker identification or verification techniques
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/04—Training, enrolment or model building
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/06—Decision making techniques; Pattern matching strategies
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/18—Artificial neural networks; Connectionist approaches
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
    • G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G10L25/24—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being the cepstrum
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination

Definitions

  • Constant Q Cepstral Coefficients are perceptually-inspired time-frequency analysis acoustic features that are found to be powerful at detecting voice spoofing attacks, namely audio playback, voice conversion and morphing, and speech synthesis attacks.
  • voice spoofing attacks namely audio playback, voice conversion and morphing, and speech synthesis attacks.
  • drawbacks of the conventional technique for obtaining CQCCs include high costs in terms of memory usage and processing time.
  • conventional systems employee CQCC features discriminate only between spoofed and non-spoofed utterances.
  • the present disclosure is directed to a methods and apparatuses for use in, for example an automatic speaker verification (ASV) system, to distinguish between a genuine, authorized speaker (e.g., a caller) and a fraudulent one, particularly one employing a replay attack, as discussed in further detail below.
  • ASV automatic speaker verification
  • a method for detecting a spoofed voice source includes receiving a voice sample, extracting at least deep acoustic features from the voice sample using a first deep neural network (DNN), and calculating, via a second DNN that receives the extracted at least deep acoustic feature, a likelihood that the voice sample includes a spoofing condition.
  • DNN deep neural network
  • the method may additionally include classifying the voice sample, using a binary classifier, as being either genuine or spoofed based on the likelihood from the second DNN.
  • the deep acoustic features are deep constant Q cepstral coefficients (CQCC).
  • the spoofing conditions include at least one of channel conditions and audio conditions.
  • the channel conditions include channel artifacts specific to at least one of different background environments, different acquisition devices, and different network infrastructures.
  • the method may also include extracting other acoustic features from the voice sample, combining the deep acoustic features with the other acoustic features to provide tandem features, and classifying the tandem features using the second DNN (X40), the second DNN configured to determine whether the tandem features include a non-spoofing condition or at least one spoofing condition. Classifying at least the deep acoustic features is part of the classifying of the tandem features.
  • the other acoustic features are sub-band cepstral coefficient (SBCC) features
  • the method may further include sub-band filtering the voice sample before extracting the other features from the filtered sample.
  • Extracting the other, SBCC features includes calculating a short-time Fourier transform (STFT) on a frame from the filtered sample, calculating a power spectrum from the STFT, calculating a log-amplitude from the power spectrum, calculating an inverse discrete cosine transform (IDCT) of the log- amplitude, and calculating dynamic features based on the IDCT.
  • STFT short-time Fourier transform
  • IDCT inverse discrete cosine transform
  • filtering the audio sample includes using a high pass filter, the filtered sample being limited to frequencies above a predetermined cutoff frequency.
  • calculating dynamic features includes calculating at least one of delta and delta-delta (acceleration) features.
  • the second DNN is configured to extract multi-class features from the at least deep acoustic features.
  • the first DNN and the second DNN each include at least an input layer, hidden layers, including one or more convolutional layers followed by a max-pooling layer, one or more fully-connected layers, and an output layer.
  • the max pooling layer of the first DNN is configured to extract bottleneck features from the deep acoustic features, the bottleneck features being features, of the at least deep acoustic features, that are sensitive to channel artifacts.
  • batch normalization is applied, for at least one of the first DNN and the second DNN, to one or more of: the input layer, the hidden layers, the one or more fully-connected layers, and the output layer.
  • the second DNN is trained via at least one of gradient boosting and back propagation.
  • the second DNN is implemented using one or more graphics processors.
  • the configuration of the second DNN results from training the second DNN with a plurality of non-spoofed and known-spoofed voice samples.
  • an apparatus for detecting a spoofed voice source includes a receiving circuit configured to receive a voice sample, a first deep neural network (DNN) configured to extract at least deep acoustic features from the voice sample, and a second DNN configured to calculate from the deep acoustic features a likelihood that the voice sample includes a spoofing condition.
  • DNN deep neural network
  • the apparatus further includes a binary classifier configured to classify the voice sample as being either genuine or spoofed based on the likelihood from the second DNN.
  • the apparatus may further include circuitry configured to extract other acoustic features from the voice sample, and a feature concatenation device configured to combine the deep acoustic features with the other acoustic features to provide tandem features, wherein the second DNN is further configured to classify the tandem features and to determine whether the tandem features include a non-spoofing condition or at least one spoofing condition, and the configuration to classify the at least the deep acoustic features is included in said configuration to classify the tandem features.
  • FIG. 1 is a block diagram illustrating an automatic speech verification (ASV) apparatus or system according to a general embodiment
  • FIG. 2 is a block diagram of a more specific implementation of the apparatus or system illustrated in FIG. 1,
  • FIGS. 3A-C are flow charts detailing methods for classifying a voice sample as genuine or spoofed, according to respective embodiments,
  • FIG. 4 illustrates an apparatus employing score fusion according to an embodiment
  • FIG. 5 illustrates an apparatus employing feature concatenation according to an embodiment
  • FIG. 6 illustrates a process for extracting sub-band cepstral coefficient features
  • FIG. 7 is a block diagram of a second deep neural network configured to classify spoofing features from tandem features as illustrated in FIG. 5, according to an embodiment.
  • FIG. 1 may depict exemplary configurations for an apparatus of the disclosure, which is done to aid in understanding the features and functionality that can be included in the housings described herein.
  • the apparatus is not restricted to the illustrated architectures or configurations, but can be implemented using a variety of alternative architectures and configurations.
  • the apparatus is described above in terms of various exemplary embodiments and implementations, it should be understood that the various features and functionality described in one or more of the individual embodiments with which they are described, but instead can be applied, alone or in some combination, to one or more of the other embodiments of the disclosure, whether or not such embodiments are described and whether or not such features are presented as being a part of a described embodiment.
  • the breadth and scope of the present disclosure, especially in any following claims, should not be limited by any of the above-described exemplary embodiments.
  • Spoofing is a fraudulent or malicious communication sent from an unknown source disguised as a source known to the receiver.
  • ASV systems are being increasingly challenged by spoofing techniques like voice conversion, speech synthesis and replay attacks.
  • replay attacks also known as "presentation attacks”— have been shown to reduce the accuracy of conventional ASV systems by a significant margin.
  • ASV Automatic Speaker Verification
  • Replay attacks are easy to generate, requiring no expertise in speech processing and/or machine learning, and very little expertise in any case. For example, with high quality recording and playback devices, it is conceivable to generate replay attacks that are nearly indistinguishable from a genuine access.
  • FIG. 1 is a block diagram illustrating an automatic speech verification (ASV) apparatus or system 100 according to a general embodiment.
  • a voice source 10 e.g., a person or, typically fraudulently, a recording of a person
  • a voice sample 20 provides a voice sample 20 over an audio channel (e.g., telephone line, wireless channel, internet channel or the like) to the ASV apparatus or system 100.
  • the voice sample is received by a feature extraction unit 1 10.
  • the feature extraction unit 1 10 extracts from the voice sample statistical information referred to as "acoustic features" 1 15 and provides the acoustic features 1 15 to a feature classification unit 120.
  • the feature classification unit 120 classifies the acoustic features 1 15 into at least one classification 125 among two or more classifications corresponding to respective spoofing conditions that can be represented in the extracted features.
  • a binary classifier 130 may compare the resulting classification with a predetermined threshold score, resulting in a determination that the voice sample or audio source is "genuine” or "fraudulent” (or similar meanings). In some embodiments, the binary classifier 130 may compare the resulting classification with another classification 135 from previously stored low-level features for a voice model corresponding to an authorized user.
  • One or more enrollment models may be generated for each authorized user at a speaker enrollment time and stored in an enrollment database 140.
  • training models 145 including known spoofed and known clean models may be stored in training database possibly collocated with the enrollment database 140.
  • the training models may be used to train the feature classification unit 120 to discriminate one or more spoofing conditions from a genuine access.
  • the enrollment models may also be used for purposes of identifying and/or verifying a speaker.
  • the term "enrollment model" here includes an initial model generated at an enrollment time, and any updates to such model based on subsequent genuine interactions with the corresponding genuine speaker.
  • a replay attack providing a recording of a speaker may, in certain low level features, include qualities found in a played-back recording that are typically not present in a (non-played back), or live, utterance.
  • recording an audio sample may consistently introduce audio artifacts related to frequency, frequency range, dynamic power range, reverberation, noise levels in particular frequency ranges, and the like, at least some of which artifacts may be imperceptible without using specialized speech processing techniques and/or equipment such as disclosed herein.
  • a genuine speaker may consistently contact the ASV system using only a limited number of channels (e.g., via a particular home phone line and/or a particular cell phone line) having specific low-level audio qualities, thus providing an enrollment model that has particular, consistent low-level features.
  • the ASV system may, as part of, instead of, or as the enrollment model, capture and store only the low-level features for later use in distinguishing spoofed and genuine access.
  • the ASV apparatus or system 100 may incorporate one or more programmed or application-specific processors, associated memory, input and output terminals, and the like, and may be implemented in a single monolithic device or distributed amongst plural proximate devices.
  • the hardware elements may incorporate devices disposed remotely and configured to communicate data using a data network.
  • the hardware may include receiving circuitry or a receiving circuit (illustrated only implicitly) such as, but not limited to, telephony or data communication circuitry configured to receive the voice sample 20.
  • the features of the ASV system 100 may include computer executable instructions on a non-transient computer-readable medium.
  • FIG. 2 is a block diagram illustrating a more specific implementation 200 of the apparatus or system 100 illustrated in FIG. 1.
  • the ASV apparatus or system 200 includes a deep neural network (DNN) 210 (corresponding to the feature extraction unit 110 in FIG. 1) configured to extract deep acoustic features 215 (corresponding to acoustic features 115).
  • the deep acoustic features 215 are received by a second deep neural network 220 (corresponding to feature classification unit 120) configured to determine (e.g., calculate) from the deep acoustic features 215 a likelihood 225 that the voice sample 20 includes a spoofing condition.
  • DNN deep neural network
  • a binary classifier 230 may compare the determined or calculated likelihood 225 (e.g., a likelihood score corresponding to classification 125) with a previously determined threshold score.
  • the binary classifier 230 may thus output an indication of "genuine” or “spoofed” depending on how the determined or calculated likelihood 225 compares with the threshold score.
  • the binary classifier 230 may compare the likelihood 225 with a previously determined likelihood 235 (corresponding to classification 135) that the spoofing condition was p resent in a previously processed enrollment sample.
  • the binary classifier 230 may employ a (different) predetermined threshold, comparing the calculated likelihood 225 (e.g., a likelihood score) is within a predetermined threshold distance or difference from the previously determined likelihood 235, the received voice sample 20 may be indicated as genuine. Otherwise, the received voice sample 20 may be indicated as spoofed.
  • the ASV apparatus or system 200 may incorporate one or more programmed or application-specific processors, associated memory, input and output terminals, and the like, and may be implemented in a single monolithic device or distributed amongst plural proximate devices.
  • the hardware elements may incorporate devices disposed remotely and configured to communicate data using a data network.
  • the hardware may include receiving circuitry or a receiving circuit (illustrated only implicitly) such as, but not limited to, telephony or data communication circuitry configured to receive the voice sample 20.
  • the features of the ASV system 100 may include computer executable instructions on a non-transient computer-readable medium.
  • the first DNN in each embodiment disclosed herein may be implemented using a convolutional neural network where an input to the convolutional neural network is a raw audio signal, and output is the desired Deep CQCC features.
  • the loss function used during training may be, but is not limited to, a mean square error between a "handcrafted" (or conventional) CQCC and a normalized CQCC.
  • the error is then used to adjust values at each node of the convolutional neural network. Iterations of the training are aimed at minimizing the error until a predetermined minimum error is reached. Batch-normalization and dropout techniques may be used, despite additional computational cost and memory required for same.
  • FIG. 3A is a flow chart detailing a method 300a for classifying a voice sample as genuine or spoofed, e.g., using the ASV 200 of FIG. 2.
  • an automated speech verification (ASV) apparatus or system receives a voice sample (e.g., 20) from a voice source (e.g., 10).
  • the voice sample may be received in a phone call from a wired or wireless telephone over a conventional or specialized telephony channel, a voice command captured directly from a voice source by an electronic, voice-responsive, intelligent personal assistant (e.g., AMAZON ECHO and the like) or voice-responsive phone app, or received by other electronic means.
  • AMAZON ECHO intelligent personal assistant
  • a genuine voice source is typically a live, preauthorized person speaking directly to the ASV via a transient-signal channel (e.g., live telephone call), whereas a "spoofed" voice source may include intentionally fraudulent recordings of a genuine voice, in the case of a replay attack, or in some cases may include unintended or accidental voice sources such as a television or radio in the background (e.g., in the case of an electronic personal assistant).
  • Other spoofed voice sources may, as suggested above, include voice conversion and speech synthesis, each of which introduces its own acoustic artifacts and/or other unique qualities distinguishable by an appropriately trained DNN.
  • DNN deep acoustic features are extracted from the voice sample. This may be accomplished via a first deep neural network (DNN) (e.g., DNN 210 in FIG. 2) that has been configured or trained for such extraction.
  • the first DNN may be supplied having been pre-trained.
  • DNN circuitry may have been particularly arranged, or DNN code may have been configured for execution, from a storage medium, by one or more processors, having already been trained by iterative processing of training samples known to include variations of low level acoustic qualities of interest. In some embodiments, however, the first DNN may be configured to be trained by such iterative processing.
  • the extracted deep acoustic features may be classified into at least one of two or more classifications corresponding to respective spoofing or channel conditions discernable in the extracted deep acoustic features by a second DNN (e.g., DNN 220 in FIG. 2).
  • the second DNN may in some embodiments be trained to discriminate between, e.g., presence and absence of spoofing condition.
  • the second DNN may be configured to discriminate between each of a plurality of spoofing conditions.
  • the second DNN may be configured to discriminate between deep acoustic features for samples having low level features corresponding respectively to a replay attack, voice conversion and speech synthesis, and to an absence thereof.
  • the second DNN may provide a likelihood score for one or more of the spoofing types.
  • the second DNN may provide a likelihood score for presence of features corresponding to a replay attack, another likelihood score for presence of features corresponding to voice conversion, a third likelihood score for presence of features corresponding to speech synthesis.
  • a high likelihood may correspond with a particular respective category.
  • Low likelihood scores for each category may then suggest absence of spoofing conditions, and thus may be categorized as "genuine".
  • the second DNN may be trained to distinguish different channel conditions. For example, the second DNN may distinguish a voice sample sourced from outdoors, from a particular room size; from a particular device, or the like.
  • a likelihood score or classification may be further classified by a binary classifier (such as binary classifier 230) as genuine or spoofed.
  • the binary classifier may be configured or trained to discriminate between the likelihood score or categorization resulting from operation S330 and a likelihood score or categorization associated with previously captured enrollment sample of an authorized user.
  • the likelihood score or categorization of the enrollment sample may be predetermined and stored in association with the enrollment sample, e.g., in an enrollment database (e.g., enrollment database 140).
  • the likelihood score or categorization for the enrollment sample may be determined when used for comparison.
  • the binary classifier may include a simple comparison of likelihood score.
  • An absolute difference between a likelihood score resulting from operation S330 and a likelihood score associated with previously captured enrollment sample may be compared with a threshold difference.
  • a threshold difference For example, an absolute difference less than or equal to the threshold difference may indicate a genuine access.
  • the threshold difference may be predetermined and stored, or may be calculated based on various factors.
  • the threshold difference may vary depending on factors such as gender, remoteness of a user at enrollment time or at test time (determined in another operation), or other factors that may affect a likelihood score.
  • users having certain vocal disabilities may routinely use a recording or speech synthesis to access an account, and may therefore require a different, more lenient threshold in order to be verified.
  • the deep acoustic features extracted in operation S320 may be Constant-Q Cepstral Coefficients (CQCCs). It is acknowledged the CQCC features may be obtained through means other than DNN. However, the inventors have recognized a benefit in cost and efficiency in utilizing a DNN for such extraction. Moreover, a DNN (such as first DNN 210) may be trained to distinguish features that would be much more difficult to account for using "handcrafted" CQCC features. For example, conventional CQCC features are typically derived from a perceptually inspired Constant-Q transform and may be classified using Gaussian Mixture Modeling (GMM).
  • GMM Gaussian Mixture Modeling
  • Spoofing conditions may include audio conditions and/or channel conditions.
  • Audio conditions may include, but are not limited to, recording artifacts, playback artifacts, and/or artifacts that result from speech synthesis, speech conversion, and/or voice distortion.
  • even very sophisticated playback equipment may introduce spectral, temporal, and/or more complex artifacts that, although not necessarily aurally perceptible, may be identified by an appropriately trained deep neural network. Less sophisticated playback equipment naturally produces even more acoustic artifacts.
  • recording equipment and/or environment may introduce acoustic artifacts such, but not limited to, noise, hum, temporal stretch, and/or frequency limitations in addition to more complex artifacts such as power spectrum inconsistencies, and the like.
  • audio conditions may include audio artifacts specific to one or more non-expected audio capture devices.
  • a genuine user may, in all previous communications, have accessed the ASV using a particular mobile telephone handset.
  • a spoofing call using a different telephone handset, a computer microphone, Internet of Things (IoT) device (e.g., AMAZON ECHO, GOOGLE HOME and the like) or other unexpected audio captured device, may include artifacts specific to a particular device, and may be different from artifacts introduced by the device employed by the genuine user at enrollment time and/or in subsequent voice communications.
  • IoT Internet of Things
  • Channel conditions may include low level artifacts introduced by the channel other than by recording or playback.
  • a voice sample transmitted over a spoofed telephone number may include artifacts not expected in a voice sample from a genuine user, such as, but not limited to quantization noise, switching noise, DC offset, filtering, and other spectral and/or temporal artifacts that may differ from those in an enrollment model, as well as higher order features that may be identified only by processing the voice sample, such as unexpected features in power spectral density and the like.
  • Such spoofing conditions may be identified by an appropriately trained deep neural network such as described above. Training of such DNN may include conventional training techniques, while focusing on discriminating between non-spoofing conditions and known spoofing conditions.
  • FIGS. 3B, 3C are flow charts depicting methods 300b, 300c for classifying a voice sample as genuine or spoofed according to other embodiments.
  • FIGS. 3B-3C illustrates use of an additional set of acoustic features extracted in parallel with the above described deep acoustic features. Operation having a same label as in FIG. 3A have function described previously.
  • the deep acoustic features and other acoustic features are extracted separately and classified separately in operations S330 and S332, e.g., using independent second and third DNNs trained to identify low level features represented in the extracted features.
  • Operations S330 and S332 may further include determining or calculating a likelihood score for each set of extracted features indicating likelihood that the extracted features include spoofing conditions.
  • the likelihood scores are "fused", e.g., according to conventional score fusion techniques.
  • the resultant fused likelihood score may be classified in operation S340 to indicate a genuine access or a spoofed access.
  • operation S340 may include comparison of the fused likelihood score with a predetermined threshold or against a likelihood score derived from a previously saved enrollment sample for the genuine user.
  • Deep acoustic features e.g., Deep CQCC
  • other acoustic features are extracted
  • the features are combined, or placed in tandem to create "tandem features" at operation S325.
  • the tandem features are classified in the same manner as described above for operation S330 and the resultant classification or likelihood score is used, at operation S340, to classify the voice sample (e.g., voice sample 20) as genuine or spoofed.
  • the elements of methods 300a-300c may be provided as computer-executable instructions stored on one or more non-transient computer readable media.
  • the elements of methods 300a-300c may be instantiated in hardware or a mixture of software/firmware and hardware as described herein.
  • the methods of FIGS. 3B-3C may be implemented as described below.
  • FIGS. 4-5 illustrate apparatuses or systems 400, 500 respectively configured to incorporate the methods in FIGS. 3B-3C.
  • the deep acoustic features corresponding to FIGS. 3B-3C may be— as shown in FIGS. 4-5— Deep CQCC features used in combination (e.g., "tandem features") with other features extracted from the voice sample 20.
  • a voice sample 20 from a voice source 10 is received by both a first DNN 410, 510 (as described above with respect to first DNN 210, e.g., configured to extract deep CQCC features) and another feature extractor 412, 512.
  • the another feature extractor 412, 512 may be configured to extract conventional features, such as Mel -frequency cepstrum coefficients (MFCCs), low frequency cepstrum coefficients (LFCC), sub-band cepstrum coefficients (SBCCs, introduced and described in detail below) or other coefficients known to provide useful low-level features for use in automated speaker verification.
  • FIG. 4 illustrates an apparatus 400 employing score fusion for likelihood scores respectively generated for each of the two sets of extracted features according to an embodiment.
  • FIG. 5 illustrates an apparatus 500 employing feature concatenation of features respectively generated by first DNN 510 and another feature extractor 512.
  • the features 415 extracted by the first DNN 410 and the features 417 extracted by the other feature extractor 412 may independently be processed, e.g., by respective second and third deep neural networks 420, 422, to generate respective likelihood scores indicating a likelihood that the voice sample 20 includes low-level feature represented in the respective extracted features (415, 417).
  • the likelihood scores may be "fused" according to conventional techniques at a score fusion unit or algorithm 425 in order for the fused scores to be considered together by a decision unit 430 (e.g., binary classifier 230 configured as described above) to indicate a genuine access or a spoofed one.
  • a decision unit 430 e.g., binary classifier 230 configured as described above
  • one or both of the likelihood scores may be adjusted according to a predetermined algorithm, in some cases weighting and/or shifting at least one of the likelihood scores in order to ensure best results.
  • the extracted deep CQCC features 515 and extracted other acoustic features 517 may be concatenated by a concatenation unit or algorithm 518 or otherwise combined before being received by a second DNN 520 (e.g., corresponding to second DNN 220 in FIG. 2) for determination of a likelihood score or categorization.
  • the likelihood score or categorization may be received by a Decision unit 530, which may correspond to a binary classifier, such as binary classifier 230 in FIG. 2.
  • a classification of genuine or spoofed may be based on a comparison of the likelihood score with a predetermined threshold and/or by comparison with a likelihood score of a previously stored enrollment sample associated with a genuine user for whom the voice sample is intended to match.
  • the apparatus or systems 400, 500 may each incorporate one or more programmed or application-specific processors, associated memory, input and output terminals, and the like, and may be implemented in a single monolithic device or distributed amongst plural proximate devices.
  • the hardware elements may incorporate devices disposed remotely and configured to communicate data using a data network.
  • the hardware may include receiving circuitry or a receiving circuit (illustrated only implicitly) such as, but not limited to, telephony or data communication circuitry configured to receive the voice sample 20.
  • SBC sub-band cepstral coefficient
  • FIG. 6 illustrates a process 600 for extracting SBCC features.
  • SBCC features may be extracted by, at operation S610, sub-band filtering the voice sample before extracting the other features from the filtered sample.
  • the voice sample is high-pass filtered, such that the extracted other features are high frequency cepstral coefficients (HFCCs).
  • HFCCs high frequency cepstral coefficients
  • the filter cutoff frequency(ies) and fall-off rate(s) may be dynamically adjusted to address different types of spoofing, e.g., where the types of spoofing occur within different frequency bands.
  • the extraction itself may commence at transform operation S620 including calculation of a short-time Fourier transform (STFT) on a frame from the filtered sample.
  • a power spectrum of the STFT is calculated at operation S630.
  • STFT short-time Fourier transform
  • a log-amplitude is calculated from the power spectrum, followed at operation S650 by calculation of an inverse discrete cosine transform (IDCT) of the log amplitude.
  • Dynamic features are calculated from the IDCT at operation S660. Calculation of the dynamic features may include calculation of delta and/or delta-delta (acceleration) features. In some instances, the delta and delta-delta features may be calculated using a window size of 1.
  • Each of the first DNN (e.g., 210 in FIG. 2) and the second DNN (e.g., 220 in FIG. 2) may include an input layer, one or more hidden layers, one or more fully connected layers, and an output layer. Any of the layers may employ batch normalization. In some embodiments, batch normalization may be applied to all layers of at least the first DNN. This mitigates overfitting.
  • the inventors provided an excerpt of raw audio (25.5 ms) to an input layer of the first DNN. An audio signal with a sampling rate of 16 kHz thus corresponded to 408 samples.
  • the hidden layers consisted of a set of convolutional layers followed by a max-pooling layer to pool over time, and then another set of fully-connected layers. The number of units of the output layer is set to match the dimension of standard (e.g., "handmade") CQCC features.
  • FIG. 7 is a block diagram of a second deep neural network (DNN) 700 such as the second DNN 520 configured to classify spoofing features from tandem features as illustrated in FIG. 5.
  • Tandem features 710 illustrate concatenation of SBCC features with Deep CQCC features.
  • FIG. 7 illustrates a Deep CQCC coefficient appended to an SBCC coefficient, the reverse may be true (i.e., SBCC appended to Deep CQCC).
  • Deep CQCC may be combined with extracted features other than SBCC.
  • the resulting tandem features are received by the DNN 700, which may include a plurality of convolutional layers 720 followed by at least one max pooling layer, one or more fully connected layers 740, and target channel conditions 750.
  • the last max pooling layer 730 of the first DNN may be configured to extract bottleneck features.
  • the bottleneck features are features from the Deep CQCC that are sensitive to audio and/or channel artifacts. Improvement results in the ability of the bottleneck features to generalize to unseen spoofing conditions. That is, presently disclosed embodiments may handle unforeseen spoofing conditions more accurately than conventional anti-spoofing approaches.
  • Training of a DNN may take several paths.
  • the second DNN (e.g., 220 in FIG. 2) may be trained via conventional approaches, with optimization done via gradient boosting and back propagation.
  • the second DNN may be trained to discriminate between multiple channel conditions.
  • the second, classifying DNN may discriminate between different background environments (e.g., home, office, restaurant, street), between different acquisition devices (e.g. laptops, smart phones, AMAZON ECHO) and/or different network infrastructures (e.g. landline, cellular, and VoIP).
  • Any of the features described herein may be implemented in software or firmware executed by a programmed processor with memory, interfaces, etc., and may provide an output that can be expressed visually (e.g., via a display indication), audibly (e.g., via a particular sound), or by an action (e.g., terminating a call, joining a live operator to a call, or requesting additional information from the user.)
  • one or both the DNNs may be implemented using a graphics processor.
  • the present invention generally relates to an apparatus for performing the operations described herein.
  • This apparatus may be specially constructed for the required purposes such as a graphics processing unit (GPU), digital signal processor (DSP), application specific integrated circuit (ASIC), field programmable gate array (FPGA) special purpose electronic circuit, or it may comprise a general-purpose computer selectively activated or reconfigured by a computer program stored in the computer.
  • GPU graphics processing unit
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • Such a computer program may be stored in a non-transient computer readable storage medium, such as, but is not limited to, any type of disk including optical disks, CD-ROMs, magneto-optical disks, read-only memories (ROMs), random access memories (RAMs), EPROMs, EEPROMs, magnetic or optical cards, integrated memory, "cloud” storage, or any type of computer readable media suitable for storing electronic instructions.
  • a non-transient computer readable storage medium such as, but is not limited to, any type of disk including optical disks, CD-ROMs, magneto-optical disks, read-only memories (ROMs), random access memories (RAMs), EPROMs, EEPROMs, magnetic or optical cards, integrated memory, "cloud” storage, or any type of computer readable media suitable for storing electronic instructions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Acoustics & Sound (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Signal Processing (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Telephonic Communication Services (AREA)

Abstract

L'invention concerne un système de vérification de locuteur automatisée (ASV) qui incorpore un premier réseau neuronal profond pour extraire des caractéristiques acoustiques profondes, telles que des caractéristiques CQCC profondes, d'un échantillon vocal reçu. Les caractéristiques acoustiques profondes sont traitées par un second réseau neuronal profond qui classifie les caractéristiques acoustiques profondes selon une probabilité déterminée qu'elles comprennent un état d'arnaque. Un classificateur binaire classe ensuite l'échantillon vocal comme étant authentique ou une arnaque.
PCT/US2018/020624 2017-03-03 2018-03-02 Procédé et appareil de détection d'états d'arnaque WO2018160943A1 (fr)

Priority Applications (4)

Application Number Priority Date Filing Date Title
AU2018226844A AU2018226844B2 (en) 2017-03-03 2018-03-02 Method and apparatus for detecting spoofing conditions
CA3054063A CA3054063A1 (fr) 2017-03-03 2018-03-02 Procede et appareil de detection d'etats d'arnaque
EP18712052.2A EP3590113B1 (fr) 2017-03-03 2018-03-02 Procédé et appareil de détection d'états d'arnaque
AU2021277642A AU2021277642B2 (en) 2017-03-03 2021-11-30 Method and apparatus for detecting spoofing conditions

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US201762466911P 2017-03-03 2017-03-03
US62/466,911 2017-03-03
US15/910,387 US10692502B2 (en) 2017-03-03 2018-03-02 Method and apparatus for detecting spoofing conditions
US15/910,387 2018-03-02

Publications (1)

Publication Number Publication Date
WO2018160943A1 true WO2018160943A1 (fr) 2018-09-07

Family

ID=63355275

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2018/020624 WO2018160943A1 (fr) 2017-03-03 2018-03-02 Procédé et appareil de détection d'états d'arnaque

Country Status (5)

Country Link
US (2) US10692502B2 (fr)
EP (1) EP3590113B1 (fr)
AU (2) AU2018226844B2 (fr)
CA (1) CA3054063A1 (fr)
WO (1) WO2018160943A1 (fr)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110827798A (zh) * 2019-11-12 2020-02-21 广州欢聊网络科技有限公司 一种音频信号处理的方法及装置
WO2021164256A1 (fr) * 2020-02-17 2021-08-26 平安科技（深圳）有限公司 Procédé, appareil et dispositif de traitement de signaux de parole
US20220172739A1 (en) * 2020-12-02 2022-06-02 Google Llc Self-Supervised Speech Representations for Fake Audio Detection

Families Citing this family (75)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10672403B2 (en) 2017-02-07 2020-06-02 Pindrop Security, Inc. Age compensation in biometric systems using time-interval, gender and age
CA3054063A1 (fr) * 2017-03-03 2018-09-07 Pindrop Security, Inc. Procede et appareil de detection d'etats d'arnaque
WO2019002831A1 (fr) 2017-06-27 2019-01-03 Cirrus Logic International Semiconductor Limited Détection d'attaque par reproduction
GB201713697D0 (en) 2017-06-28 2017-10-11 Cirrus Logic Int Semiconductor Ltd Magnetic detection of replay attack
GB2563953A (en) 2017-06-28 2019-01-02 Cirrus Logic Int Semiconductor Ltd Detection of replay attack
GB201801530D0 (en) 2017-07-07 2018-03-14 Cirrus Logic Int Semiconductor Ltd Methods, apparatus and systems for authentication
GB201801526D0 (en) 2017-07-07 2018-03-14 Cirrus Logic Int Semiconductor Ltd Methods, apparatus and systems for authentication
GB201801532D0 (en) 2017-07-07 2018-03-14 Cirrus Logic Int Semiconductor Ltd Methods, apparatus and systems for audio playback
GB201801528D0 (en) 2017-07-07 2018-03-14 Cirrus Logic Int Semiconductor Ltd Method, apparatus and systems for biometric processes
GB201801527D0 (en) 2017-07-07 2018-03-14 Cirrus Logic Int Semiconductor Ltd Method, apparatus and systems for biometric processes
US10896673B1 (en) 2017-09-21 2021-01-19 Wells Fargo Bank, N.A. Authentication of impaired voices
GB201801874D0 (en) 2017-10-13 2018-03-21 Cirrus Logic Int Semiconductor Ltd Improving robustness of speech processing system against ultrasound and dolphin attacks
GB201801664D0 (en) 2017-10-13 2018-03-21 Cirrus Logic Int Semiconductor Ltd Detection of liveness
GB201801663D0 (en) 2017-10-13 2018-03-21 Cirrus Logic Int Semiconductor Ltd Detection of liveness
GB201804843D0 (en) 2017-11-14 2018-05-09 Cirrus Logic Int Semiconductor Ltd Detection of replay attack
GB2567503A (en) 2017-10-13 2019-04-17 Cirrus Logic Int Semiconductor Ltd Analysing speech signals
GB201801661D0 (en) 2017-10-13 2018-03-21 Cirrus Logic International Uk Ltd Detection of liveness
GB201803570D0 (en) 2017-10-13 2018-04-18 Cirrus Logic Int Semiconductor Ltd Detection of replay attack
GB201719734D0 (en) * 2017-10-30 2018-01-10 Cirrus Logic Int Semiconductor Ltd Speaker identification
GB201801659D0 (en) * 2017-11-14 2018-03-21 Cirrus Logic Int Semiconductor Ltd Detection of loudspeaker playback
US10657971B1 (en) * 2017-12-15 2020-05-19 NortonLifeLock Inc. Systems and methods for detecting suspicious voice calls
US11475899B2 (en) 2018-01-23 2022-10-18 Cirrus Logic, Inc. Speaker identification
US11264037B2 (en) 2018-01-23 2022-03-01 Cirrus Logic, Inc. Speaker identification
US11735189B2 (en) 2018-01-23 2023-08-22 Cirrus Logic, Inc. Speaker identification
KR102531654B1 (ko) * 2018-05-04 2023-05-11 삼성전자주식회사 음성 입력 인증 디바이스 및 그 방법
US10529356B2 (en) 2018-05-15 2020-01-07 Cirrus Logic, Inc. Detecting unwanted audio signal components by comparing signals processed with differing linearity
US10692490B2 (en) 2018-07-31 2020-06-23 Cirrus Logic, Inc. Detection of replay attack
US10915614B2 (en) 2018-08-31 2021-02-09 Cirrus Logic, Inc. Biometric authentication
US11037574B2 (en) 2018-09-05 2021-06-15 Cirrus Logic, Inc. Speaker recognition and speaker change detection
CN109377983A (zh) * 2018-10-18 2019-02-22 深圳壹账通智能科技有限公司 一种基于语音交互的骚扰电话拦截方法及相关设备
CN109300479A (zh) * 2018-10-31 2019-02-01 桂林电子科技大学 一种回放语音的声纹识别方法、装置及存储介质
CN109243487B (zh) * 2018-11-30 2022-12-27 宁波大学 一种归一化常q倒谱特征的回放语音检测方法
CN109801638B (zh) * 2019-01-24 2023-10-13 平安科技（深圳）有限公司 语音验证方法、装置、计算机设备及存储介质
US11289098B2 (en) * 2019-03-08 2022-03-29 Samsung Electronics Co., Ltd. Method and apparatus with speaker recognition registration
US12015637B2 (en) 2019-04-08 2024-06-18 Pindrop Security, Inc. Systems and methods for end-to-end architectures for voice spoofing detection
CN110298150B (zh) * 2019-05-29 2021-11-26 上海拍拍贷金融信息服务有限公司 一种基于语音识别的身份验证方法及系统
CN110232928B (zh) * 2019-06-13 2021-05-25 思必驰科技股份有限公司 文本无关说话人验证方法和装置
CN110232927B (zh) * 2019-06-13 2021-08-13 思必驰科技股份有限公司 说话人验证反欺骗方法和装置
CN110223676A (zh) * 2019-06-14 2019-09-10 苏州思必驰信息科技有限公司 欺骗录音检测神经网络模型的优化方法及系统
CN110211604A (zh) * 2019-06-17 2019-09-06 广东技术师范大学 一种用于语音变形检测的深度残差网络结构
JP7396376B2 (ja) * 2019-06-28 2023-12-12 日本電気株式会社 なりすまし検出装置、なりすまし検出方法、及びプログラム
CN110491391B (zh) * 2019-07-02 2021-09-17 厦门大学 一种基于深度神经网络的欺骗语音检测方法
CN112447187B (zh) * 2019-09-02 2024-09-06 富士通株式会社 声音事件的识别装置和方法
US11158329B2 (en) * 2019-09-11 2021-10-26 Artificial Intelligence Foundation, Inc. Identification of fake audio content
CN110689885B (zh) * 2019-09-18 2023-05-23 平安科技（深圳）有限公司 机器合成语音识别方法、装置、存储介质及电子设备
CN110797031A (zh) * 2019-09-19 2020-02-14 厦门快商通科技股份有限公司 语音变音检测方法、系统、移动终端及存储介质
CN112598107A (zh) * 2019-10-01 2021-04-02 创鑫智慧股份有限公司 数据处理系统及其数据处理方法
US11039205B2 (en) 2019-10-09 2021-06-15 Sony Interactive Entertainment Inc. Fake video detection using block chain
EP4049174A4 (fr) * 2019-10-21 2024-01-03 Sony Interactive Entertainment Inc. Détection de fausse vidéo
US11551474B2 (en) 2019-10-21 2023-01-10 Sony Interactive Entertainment Inc. Fake video detection
CN110782877A (zh) * 2019-11-19 2020-02-11 合肥工业大学 基于Fisher混合特征和神经网络的语音鉴别方法和系统
WO2021126444A1 (fr) * 2019-12-20 2021-06-24 Eduworks Corporation Détection d'hameçonnage vocal en temps réel
WO2021137754A1 (fr) * 2019-12-31 2021-07-08 National University Of Singapore Conversion vocale commandée par rétroaction
CN111243621A (zh) * 2020-01-14 2020-06-05 四川大学 一种用于合成语音检测的gru-svm深度学习模型的构造方法
EP4097717A4 (fr) * 2020-01-27 2024-02-21 Pindrop Security, Inc. Système de détection de mystification robuste utilisant des réseaux neuronaux résiduels profonds
US11527245B2 (en) * 2020-04-29 2022-12-13 Rovi Guides, Inc. Systems and methods for avoiding inadvertently triggering a voice assistant
CN111835784B (zh) * 2020-07-22 2022-05-13 思必驰科技股份有限公司 用于重放攻击检测系统的数据泛化方法及系统
CN112102808A (zh) * 2020-08-25 2020-12-18 上海红阵信息科技有限公司 用于伪造语音的深度神经网络的构建方法及系统
EP4229626A4 (fr) * 2020-10-16 2024-10-09 Pindrop Security Inc Détection d'hypertrucage audiovisuel
CN112735381B (zh) * 2020-12-29 2022-09-27 四川虹微技术有限公司 一种模型更新方法及装置
CN112927694B (zh) * 2021-03-08 2022-09-13 中国地质大学（武汉） 一种基于融合声纹特征的语音指令合法性判别方法
US11483427B1 (en) 2021-04-28 2022-10-25 Zoom Video Communications, Inc. Call recording authentication
CN113436646B (zh) * 2021-06-10 2022-09-23 杭州电子科技大学 一种采用联合特征与随机森林的伪装语音检测方法
CN113646833A (zh) * 2021-07-14 2021-11-12 东莞理工学院 语音对抗样本检测方法、装置、设备及计算机可读存储介质
CN113284508B (zh) * 2021-07-21 2021-11-09 中国科学院自动化研究所 基于层级区分的生成音频检测系统
CN113299315B (zh) * 2021-07-27 2021-10-15 中国科学院自动化研究所 一种无需原始数据存储的持续性学习生成语音特征的方法
CN113314148B (zh) * 2021-07-29 2021-11-09 中国科学院自动化研究所 基于原始波形的轻量级神经网络生成语音鉴别方法和系统
CN113488070B (zh) * 2021-09-08 2021-11-16 中国科学院自动化研究所 篡改音频的检测方法、装置、电子设备及存储介质
CN113488027A (zh) * 2021-09-08 2021-10-08 中国科学院自动化研究所 一种层级分类的生成音频溯源方法及存储介质、计算机设备
CN113851147A (zh) * 2021-10-19 2021-12-28 北京百度网讯科技有限公司 音频识别方法、音频识别模型训练方法、装置、电子设备
US20240005911A1 (en) * 2022-05-27 2024-01-04 Accenture Global Solutions Limited Systems and methods to improve trust in conversations with deep learning models
CN115083423B (zh) * 2022-07-21 2022-11-15 中国科学院自动化研究所 语音鉴别的数据处理方法和装置
KR20240088457A (ko) * 2022-12-13 2024-06-20 삼성전자주식회사 합성 음성을 식별하는 전자 장치 및 그 제어 방법
CN117393000B (zh) * 2023-11-09 2024-04-16 南京邮电大学 一种基于神经网络和特征融合的合成语音检测方法
CN118038890B (zh) * 2024-04-11 2024-06-18 西北工业大学 基于特征融合和联合神经网络的细粒度水声目标识别方法

Family Cites Families (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100632400B1 (ko) 2005-11-11 2006-10-11 한국전자통신연구원 음성 인식을 이용한 입출력 장치 및 그 방법
US9767806B2 (en) 2013-09-24 2017-09-19 Cirrus Logic International Semiconductor Ltd. Anti-spoofing
US20150112682A1 (en) 2008-12-10 2015-04-23 Agnitio Sl Method for verifying the identity of a speaker and related computer readable medium and computer
US8145562B2 (en) 2009-03-09 2012-03-27 Moshe Wasserblat Apparatus and method for fraud prevention
KR101070527B1 (ko) 2009-04-24 2011-10-05 서울대학교산학협력단 빛 퍼짐을 이용한 상호작용깊이 측정장치, 측정방법 및 이를 이용한 양전자 방출 단층촬영장치
CA2716266C (fr) * 2009-10-01 2016-08-16 Crim (Centre De Recherche Informatique De Montreal) Detection de polycopie magnetique a base de contenu
US9502038B2 (en) 2013-01-28 2016-11-22 Tencent Technology (Shenzhen) Company Limited Method and device for voiceprint recognition
US9372976B2 (en) 2013-03-20 2016-06-21 Dror Bukai Automatic learning multi-modal fraud prevention (LMFP) system
US20160293167A1 (en) 2013-10-10 2016-10-06 Google Inc. Speaker recognition using neural networks
US9412365B2 (en) 2014-03-24 2016-08-09 Google Inc. Enhanced maximum entropy models
US10476872B2 (en) 2015-02-20 2019-11-12 Sri International Joint speaker authentication and key phrase identification
US11823658B2 (en) * 2015-02-20 2023-11-21 Sri International Trial-based calibration for audio-based identification, recognition, and detection system
US9721559B2 (en) * 2015-04-17 2017-08-01 International Business Machines Corporation Data augmentation method based on stochastic feature mapping for automatic speech recognition
KR102446392B1 (ko) 2015-09-23 2022-09-23 삼성전자주식회사 음성 인식이 가능한 전자 장치 및 방법
CN105450965B (zh) 2015-12-09 2019-07-19 北京小鸟看看科技有限公司 一种视频转换方法、装置和系统
US10366687B2 (en) 2015-12-10 2019-07-30 Nuance Communications, Inc. System and methods for adapting neural network acoustic models
US10235994B2 (en) * 2016-03-04 2019-03-19 Microsoft Technology Licensing, Llc Modular deep learning model
US10460747B2 (en) * 2016-05-10 2019-10-29 Google Llc Frequency based audio analysis using neural networks
US9824692B1 (en) * 2016-09-12 2017-11-21 Pindrop Security, Inc. End-to-end speaker recognition using deep neural network
CA3179080A1 (fr) * 2016-09-19 2018-03-22 Pindrop Security, Inc. Caracteristiques de bas niveau de compensation de canal pour la reconnaissance de locuteur
US10360494B2 (en) * 2016-11-30 2019-07-23 Altumview Systems Inc. Convolutional neural network (CNN) system based on resolution-limited small-scale CNN modules
US10140980B2 (en) * 2016-12-21 2018-11-27 Google LCC Complex linear projection for acoustic modeling
CA3054063A1 (fr) * 2017-03-03 2018-09-07 Pindrop Security, Inc. Procede et appareil de detection d'etats d'arnaque
CN107068167A (zh) * 2017-03-13 2017-08-18 广东顺德中山大学卡内基梅隆大学国际联合研究院 融合多种端到端神经网络结构的说话人感冒症状识别方法
US10311872B2 (en) * 2017-07-25 2019-06-04 Google Llc Utterance classifier
US10963781B2 (en) * 2017-08-14 2021-03-30 Microsoft Technology Licensing, Llc Classification of audio segments using a classification network
US10418957B1 (en) * 2018-06-29 2019-09-17 Amazon Technologies, Inc. Audio event detection
US11257503B1 (en) * 2021-03-10 2022-02-22 Vikram Ramesh Lakkavalli Speaker recognition using domain independent embedding

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
MASSIMILIANO TODISCO ET AL: "A New Feature for Automatic Speaker Verification Anti-Spoofing: Constant Q Cepstral Coefficients", ODYSSEY 2016, vol. 2016, 21 June 2016 (2016-06-21), pages 283 - 290, XP055476497, ISSN: 2312-2846, DOI: 10.21437/Odyssey.2016-41 *
MD JAHANGIR ALAM ET AL: "Spoofing Detection on the ASVspoof2015 Challenge Corpus Employing Deep Neural Networks", ODYSSEY 2016, vol. 2016, 21 June 2016 (2016-06-21), pages 270 - 276, XP055476292, ISSN: 2312-2846, DOI: 10.21437/Odyssey.2016-39 *
NAGARSHETH ET AL.: "Replay Attack Detection using DNN for Channel Discrimination", INTERSPEECH, August 2017 (2017-08-01)
PARAV NAGARSHETH ET AL: "Replay Attack Detection Using DNN for Channel Discrimination", INTERSPEECH 2017, 20 August 2017 (2017-08-20), ISCA, pages 97 - 101, XP055476190, DOI: 10.21437/Interspeech.2017-1377 *
TODISCO ET AL.: "A New Feature for Automatic Speaker Verification Anti-Spoofing: Constant Q Cepstral Coefficients", ODYSSEY, 2016
WU ZHIZHENG ET AL: "ASVspoof: The Automatic Speaker Verification Spoofing and Countermeasures Challenge", IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, IEEE, US, vol. 11, no. 4, 17 February 2017 (2017-02-17), pages 588 - 604, XP011649191, ISSN: 1932-4553, [retrieved on 20170512], DOI: 10.1109/JSTSP.2017.2671435 *
ZHANG CHUNLEI ET AL: "An Investigation of Deep-Learning Frameworks for Speaker Verification Antispoofing", IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, IEEE, US, vol. 11, no. 4, 16 January 2017 (2017-01-16), pages 684 - 694, XP011649474, ISSN: 1932-4553, [retrieved on 20170515], DOI: 10.1109/JSTSP.2016.2647199 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110827798A (zh) * 2019-11-12 2020-02-21 广州欢聊网络科技有限公司 一种音频信号处理的方法及装置
WO2021164256A1 (fr) * 2020-02-17 2021-08-26 平安科技（深圳）有限公司 Procédé, appareil et dispositif de traitement de signaux de parole
US20220172739A1 (en) * 2020-12-02 2022-06-02 Google Llc Self-Supervised Speech Representations for Fake Audio Detection
WO2022119699A1 (fr) * 2020-12-02 2022-06-09 Google Llc Détection de faux audio
US11756572B2 (en) 2020-12-02 2023-09-12 Google Llc Self-supervised speech representations for fake audio detection
US20230386506A1 (en) * 2020-12-02 2023-11-30 Google Llc Self-supervised speech representations for fake audio detection

Also Published As

Publication number Publication date
US11488605B2 (en) 2022-11-01
AU2021277642A1 (en) 2021-12-23
US20200321009A1 (en) 2020-10-08
AU2018226844B2 (en) 2021-11-18
EP3590113A1 (fr) 2020-01-08
AU2021277642B2 (en) 2023-06-15
EP3590113B1 (fr) 2024-05-29
US10692502B2 (en) 2020-06-23
AU2018226844A1 (en) 2019-09-19
CA3054063A1 (fr) 2018-09-07
US20180254046A1 (en) 2018-09-06

Similar Documents

Publication Publication Date Title
AU2021277642B2 (en) Method and apparatus for detecting spoofing conditions
US11657823B2 (en) Channel-compensated low-level features for speaker recognition
US11823679B2 (en) Method and system of audio false keyphrase rejection using speaker recognition
US11670304B2 (en) Speaker recognition in the call center
US11948553B2 (en) Systems and methods of speaker-independent embedding for identification and verification from audio
US20120290297A1 (en) Speaker Liveness Detection
CN116490920A (zh) 用于针对由自动语音识别系统处理的语音输入检测音频对抗性攻击的方法、对应的设备、计算机程序产品和计算机可读载体介质
EP3516652B1 (fr) Caractéristiques de bas niveau à compensation de canal pour la reconnaissance du locuteur
US20240363100A1 (en) Active voice liveness detection system
US20240363125A1 (en) Active voice liveness detection system
Wickramasinghe Replay detection in voice biometrics: an investigation of adaptive and non-adaptive front-ends
WO2023149998A1 (fr) Détection de parole synthétique à l'aide d'un modèle adapté à des données audio d'interlocuteur
CN116631401A (zh) 一种语音通话录音数据处理方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18712052

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 3054063

Country of ref document: CA

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2018226844

Country of ref document: AU

Date of ref document: 20180302

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 2018712052

Country of ref document: EP